Microsoft TRELLIS.2 on Apple Silicon, No Nvidia GPU
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Microsoft released a PyTorch‑MPS port of TRELLIS.2 that replaces all CUDA calls, enabling inference on macOS Apple Silicon (M1 or later) without Nvidia hardware.
- Apple’s M4 Pro can generate a 400 K‑vertex mesh from a single image in roughly 3.5 minutes, using about 18 GB of its 24 GB unified memory.
- Nvidia’s CUDA‑only libraries were previously required for TRELLIS.2, but the new port removes that dependency, potentially reducing demand for Nvidia GPUs in 3D workloads.
- TRELLIS.2 now runs with pure‑Python sparse convolution and SDPA attention, which is about ten times slower than the original CUDA kernels, highlighting a performance trade‑off.
- HuggingFace hosts the gated model weights (dinov3‑vitl16, RMBG‑2.0) needed for the Mac port, requiring users to log in and request instant access.
Why it matters: Apple developers and 3D artists gain native access to Microsoft’s TRELLIS.2 on M‑series Macs, cutting the need for Nvidia GPUs and potentially shrinking Nvidia’s hardware revenue share by an estimated single‑digit percent.
Ask SkimNews
